Automated Pole-Selection: Proof-of-Concept & Validation

Jeroen Lanslots, Bert Rodiers, Bart Peeters

LMS International, TST Technology Deployment & Research
Researchpark Z1, Interleuvenlaan 68, B-3001, Leuven, Belgium
e-mail: jeroen.lanslots@lms.be

Abstract
System identification from measured MIMO data plays a crucial role in structural dynamics and vibro-
acoustic system optimization. The most popular modelling approach is based on the modal analysis con-
cept, leading to an interpretation in terms of eigenmodes, visualized in so-called stabilization diagrams.
Typically, the number of modes is very high (often over 100), including modes with high damping and high
modal overlap. Therefore, a key problem of the system identification process is the selection of the correct
model order and related to this, the selection of valid system poles. So far, this has been a problem to be
solved by human expert engineers. This paper presents a proof-of-concept of a completely automated modal
analysis process. In doing so, this paper presents a benchmark test that will proof that the performance of the
automated tool is equal to or better than the performance of human expert engineers and as such will receive
confidence for usage.

1 Introduction

System identification [1] is a standard tool for the analysis of artificially or ambient excited vibrating struc-
tures. In modal analysis, a number of parameter estimation techniques such as LSCE and PolyMAX are used
to estimate the parameters that identify a system correctly (Section 3). These so-called modal parameters
are resonance frequencies, damping ratios and modal shapes. To reduce the bias on the estimates and allow
the model to capture all relevant characteristics of the structure, the identification order is usually chosen
quite high. However, the higher the model order is chosen, the higher number of estimated poles will be
calculated. This results in the occurrence of so-called spurious or mathematical poles that do not represent
the physical behavior of the structure.
The results of these calculations are presented in a so-called stabilization diagram from which stabilized
modes can be picked. Such a diagram shows the evolution of frequency, damping, and mode/participation
vectors as the number of modes is increased. The physical modes can then be seen as those modes for which
the frequency, damping, and mode/participation vector values do not change significantly. This is exactly the
task of assessing a stabilization diagram, and so far experienced engineers do this by hand.
To automate this, an intelligent rule-based technique was developed that analyzes the stabilization diagram
by selecting only the physical poles. This technique is based on knowledge acquisition that is achieved
by observing skilled engineers. It has the advantage that it is much faster, that less-experienced people
have access to knowledge from expert engineers, and furthermore, the method is deterministic and does not
depend on the parameter estimation method used to obtain the stabilization diagram.
With the latest generation parameter estimation methods, stabilization diagrams have been cleaned of a lot
of the spurious poles. This opens the opportunity for automated pole selection methods to be included as the
final step, hereby completely automating the modal analysis process. In section 4, a proof of concept will be
given by a combination of standard LMS Test.Lab software and Windows Automation.
Section 6 reports on a benchmark test that seeks to determine the performance of the automated rule-base
intelligence method. In doing so, three groups have been created:

expert engineers using implicit knowledge: hence expertise, and experience;

non-expert engineers using explicit knowledge: they will be using the rules as they have been extracted
in the knowledge acquisition process;

the automated pole-selection method: using the same extracted rules.

To compare the quality of the resulting pole selections, a number of validation criteria, presented in section
5, will be used. Section 7 will present conclusions on the performance of the automated tool with respect to
the performance of human novice and expert engineers.

2 Overview of existing pole-selection methods

With the increasing use of modal analysis as a standard tool by many, also less experienced, users, the strong
need is expressed to automate the process. Researched solutions include estimation methods that are much
more robust with respect to the appearance of spurious poles such as PolyMAX. Nevertheless, the route to
automation still requires discrimination methods to distinguish physical from mathematical poles.
Probably, the most natural way is trying to capture the decisions that an experienced modal analyst takes,
based on a stabilization diagram, by rules that can be implemented as an autonomous procedure. Examples
of such rules can be found in [2, 3, 4]. The rule-based approach [2, 3] will be used for the benchmark test,
and is further described in section 4.
The amount of pole information that can be provided in a stabilization diagram is in fact bounded by the
ability of the human mind to interpret them at a glance. Automated procedures do not suffer from this con-
straint and the pole classification can be based on much more information. Therefore research is performed
on additional criteria that describe the stability of a pole. A first set of criteria is generic (i.e. independent
of the system identification technique) and looks for instance at the complexity of the modal vector [5, 6, 7].
As mathematical poles are often characterized by high complexity, this criterion is valid if real modes are
expected (loosely, if the distribution of the damping mechanisms along the vibrating structure is similar to its
mass and stiffness distribution). Another criterion to get confidence in a pole of an identified MIMO model
is to verify whether the pole is still present if different SIMO or even SISO models are identified by selecting
individual inputs or outputs from the data [3, 6, 8, 9].
Other pole validation criteria emerged in combination with specific system identification techniques. For
instance in [10], the autonomous procedure heavily relies upon the Consistent-Mode Indicator which was
presented as a byproduct of the Eigensystem Realization Algorithm (ERA) [11]. Based on the realization
theory, ERA identifies a state-space model from impulse response functions. Very interesting are the stochas-
tic pole validation criteria that were formulated in combination with frequency-domain maximum likelihood
estimation [6, 7, 12, 13]. These stochastic criteria propagate the data uncertainty to features such as: uncer-
tainty on pole estimate (large uncertainty may indicate a mathematical pole), pole-zero pairs (large amount
of zeros within a certain confidence circle around a pole is an indication of a mathematical pole), pole-zero
correlation coefficient (high correlation again indicates a non-physical origin of pole).
A final set of criteria was developed to operate in combination with subspace identification [3, 9, 14]: modal
model reduction (the effect of removing a mode from the state-space model is studied, where a small effect
may indicate a mathematical pole), relation between balanced and modal form (the energy interpretation of
the states of a balanced model is heuristically transferred to the poles of the system, where low-energy
poles are considered as non-physical), forced pole-zero cancellations (if the effect on the model of shifting a
pole towards its closest zeros is limited, this was probably a mathematical pole).
Evidently none of the methods works in all cases: many of them may give unclear separation in case of
low signal-to-noise ratios; some of them may fail to label a local or weakly excited pole as a physical one;
the mode complexity indicators fail if the structure exhibits complex physical modes. However, combining
multiple criteria by clustering techniques drastically improves the pole-discrimination capacity. For instance,
in [6, 13] Fuzzy C-means clustering is used for this purpose: the result is a membership function value for
each pole that indicates whether it belongs to the class of physical or mathematical poles.
Clustering techniques are also used to automatically interpret a classical stabilization diagram [15]. In such
an approach, advanced extensions of the original Fuzzy C-means clustering technique are used to cluster
poles in terms of frequency and damping. Based on the final cluster centers, the most stable poles are
chosen. In addition, [15] presents Genetic Algorithms as a suitable technique that can be used to initialize
the fuzzy clustering and/or perform a stand-alone clustering.

3 Parameter estimation methods

Because of the complexity of the structural dynamics identification problem, a large variety of dedicated
modal parameter estimation methods were proposed over the years. In most methods, the raw input-output
data are first processed into a non-parametric system description matrix, consisting of FRFs (Frequency
Response Functions) or IRs (Impulse Responses), resulting in a significant data reduction and allowing the
use and combination of non-simultaneously measured data. This also fits the test reality where the modal
testing process is often separated in time (and location) from the actual analysis process.
The matrix of FRFs or IRs can then be used as input data to establish the parameters of a system model
such as a frequency or time domain state-space model (referred to as Eigensystem Realization (ERA) [11] or
Subspace Identification [16], due to the inherent SVD model reduction), or to directly identify the structural
parameters of a model described by the constitutive equations (Direct Parameter Estimation [17]). Both ap-
proaches have been implemented in the time as well as in the frequency domain in many different algorithms.
Alternatively, model formulations can be used which directly exploit specific properties of the FRFs (or IRs).
For example, the FRF and IR matrices can be expressed directly in terms of the modal parameters:
Nm 


Qr {}r {}t r Q {}r {}t
[H(j)] = + r r
(1)
(j r ) (j r )
r=1
Nm 


r t
[h(t)] = Qr {}r {}tr er t + Qr {}r {}t
r e (2)
r=1

where Qr : modal scaling factor, {}r : modal vector r, r : system pole r + jr , r : damping factor, r :
damped natural frequency, [V ] := [{}1 . . . {}N {}1 . . . {}Nm ]: modal vector matrix.
Directly solving equation (1) leads to the non-linear frequency domain method [18], which, when used in its
original form, requires a large non-linear set of equations to be solved in an iterative way. This proves to be
unpractical for most modal data sets hence alternative methods have been developed. Below, two prominent
ones are presented. The discussed methods are applicable to both FRFs (input-output data) and output-only
spectra (operational data).

3.1 LSCE

The still most widely used approach is the Polyreference Least Squares Complex Exponential (LSCE)
method which dates from the early 80s [19]. The method starts from the Impulse Responses (IR) between
the measured inputs and outputs and yields global estimates of poles and the modal participation factors.
Define the impulse response function matrix [hk ] at time instant k between Nresp responses and Nref inputs.
Mathematically, the Polyreference LSCE will decompose the correlation functions as a sum of decaying
sinusoids. So, N m
[hk ] = r kt {L}T + {} er kt {L}T or
r=1 {}r e r r r
(3)
N m k {L}T + {} k {L}T
[hk ] = r=1 {}
r r r r r r

where system pole r = er t and {L}r is a column vector of Nref modal participation factors, which are
multipliers which are constant for all response stations for the r-th mode. The combinations of complex ex-

ponential and constant multipliers, r {L}Tr or r {L}Tr are a solution of the following matrix finite difference
equation of order t:

kr {L}Tr [I] + k1
r {L}Tr [W ]1 + . . . + rkt {L}Tr [W ]t = {0}T (4)

where [W ]1 . . . [W ]t are coefficient matrices with dimension Nref Nref . In case the system has Nm physical
modes, the order t in equation (4) has to be greater than or equal to 2Nm /Nref .
Since the impulse response functions are a linear combination of the characteristic solutions of equation (4),

r {L}Tr or r {L}Tr , they are also a solution of that equation. Hence,

[hk ][I] + [hk1 ][W ]1 + . . . + [hkt ][W ]t = [0] (5)

Equation (5) uses all response stations simultaneously enabling a global least squares estimate of the coef-
ficient matrices [W ]1 . . . [W ]t . Overdetermination is achieved by considering all available or selected time
intervals. Once the coefficient matrices are known, equation (4) can be reformulated into a generalized
Eigenvalue problem resulting into Nref t Eigenvalues r , as estimates for the system poles r and the corre-
sponding left eigenvectors {L}Tr .
As the Polyreference LSCE does not yield the mode shapes, a second step is needed to extract these using
the identified modal frequencies, damping ratios and participation factors as known parameters in equation
(1), which now becomes a simple linear set of equations.

3.2 PolyMAX

As stated before, the highly interactive pole-selection requires a large user experience.A major research topic
has hence been the development of system identification algorithms with improved stabilization behavior.
Impressive results are for example obtained with the recently developed Least Squares Complex Frequency
Domain (LSCF) method [20, 21]. This discrete frequency domain method uses a least squares or total
least squares approach to fit a rational fraction polynomial model to a MIMO FRF matrix. Originally a
common-denominator model structure was used. Recently a right matrix-fraction variant of the method was
introduced. In this so-called PolyMAX method, following model is identified from the FRF data:
p
 p 1



r r
[H()] = z [r ] z [r ] (6)
r=0 r=0

where [r ] Clm are the numerator matrix polynomial coefficients, [r ] Cmm are the denominator
matrix coefficients, and p is the model order. Note that a so-called z-domain model (i.e. frequency-domain
model that is derived from a discrete-time model) is used in (6), with:

z = ejt (7)

where t is the sampling time. Equation (6) can be written down for all values of the frequency axis of the
FRF data. Basically, the unknown model coefficients [r ], [r ] are then found as the least-squares solution
of these equations (after linearization). More details can be found in [22, 23].
Very high system orders (over 50) are clearly identified in a single-step procedure, leading to extremely clear
stabilization diagrams, and hence drastically improving the quality and the interpretability of the result. It is
a feature of the PolyMAX identification method to estimate the mathematical poles with negative damping
ratio. Hence these poles are readily excluded before constructing the stabilization diagram. The fact that
inherently unstable models are identified is not a problem as a new model is recomposed after selection of
the stable poles from the stabilization diagram.

4 Proof-of-Concept: The Modal Analysis Chain

In [2, 3], a rule-based approach is presented that identifies the steps and the knowledge used by an expert
engineer assessing a stabilization diagram. This implicit knowledge was then made explicit by casting it into
rules. In short, the engineer visually inspects the symbols in the stabilization diagrams, which are based on
similarity in frequency, damping ratio and/or mode vectors between poles belonging to subsequent model
orders. Based on these sources of information, the engineer selects a number of poles by hand. In a later
step, this pole selection is checked with validation criteria (Section 5).

Figure 1: Screenshot AASD: Automatic Assessment of Stabilization Diagrams.

The original tool that was created from this, AASD (Automatic Assessment of Stabilization Diagrams),
allows to select a stabilization diagram file that was exported from LMS Cada-X software, perform the
automatic pole-selection, and finally import this selection back into Cada-X (figure 1).
The successor of Cada-X, LMS Test.LabTM provides valuable technology that made it possible to fully
integrate the AASD tool in the Modal Analysis application. Using Windows Automation it is possible for
external applications to communicate and interact with Test.Lab. The AASD tool now grabs a generated
stabilization diagram from the Test.Lab Modal Analysis application and performs the automatic assessment.
When finished, it passes the resulting pole-selection back to Test.Lab, as if the selection was done by hand
(figure 2). The user can then assess the resulting mode shapes to his/her own judgement, and proceed to the
validation step.

Figure 2: Test.Lab Modal Analysis: PolyMAX Stabilization sheet. Pole selection by AASD.

5 Validation Criteria

A number of validation criteria exist that can be used to determine the quality of a parameter estimation of
mode shapes. The criteria presented below will be used in the benchmark test, section 6.

Synthesized vs. Measured FRFs/spectra: Test.Lab uses the selected modes to synthesize the mea-
sured FRFs/spectra. The higher the correlation with the corresponding measured FRFs/spectra, the
better the modal model. A criterion that can be used for this is the Complex Mode Indicator Function
(CMIF), where the FRFs/spectra are compressed by using their 3 principal components.

Modal Phase Collinearity (MPC): The MPC is an indicator that checks the degree of complexity
of a mode [5]. It evaluates the functional linear relation between the real and imaginary parts of the
mode shape coefficients. In general, for true physical modes the MPC value approaches 1, whereas
low values indicate again noisy or computational modes.

Mean Phase Deviation (MPD): The MPD is another statistical indicator of the complexity of a mode
shape [5]. The MPD expresses the phase scatter of each mode shape. For true physical mode shapes,
its value is 0.
6 Benchmark Test

The goal of the benchmark test is to give a qualitative assessment of the AASD tool, and to place it within
the spectrum of novice to experienced modal analysts. A group of 8 people was selected for this, containing
4 novices, and 4 experts, all with an engineering background. Two data sets were selected, both analyzed
with two parameter estimation methods, leading in total to four stabilization diagrams (Figures 3, 4, and 5).
Note the clean PolyMAX stabilization diagrams. The novices received a short description of the task they
had to carry out:

1. The stabilization diagram shows poles solutions of a mathematical problem at different

model orders. The task of the engineer is to select an unknown number of poles at
different frequencies. Occurance of a pole at the same frequency at increasing model
orders gives the engineer information on the physical correspondence of this pole.
2. First, the engineer will look for a vertical column of poles. Especially if this column will
contain lots of s-type (and d-type) poles. It is not important that this column should exist
at the lower model orders, nor that it is a straight column at the lower orders.
3. Next the engineer inspects the s-type and d-type poles in particularly, in which he values
the s-type poles more than the d-type poles. The engineer searches for the pole in the
column that is the most stable in frequency and which stabilizes in damping and scatter at
a certain order.

Body LSCE
65
o s f d fd d f o dds d f ff d f s ds f f f f s f f d sv f dv ff d f vf f f f f f s s d fv v sf fffv ff f
fd f f f d d s ff f sf v d
f f d v vf s f d v s ff f sf df ff dv df d f f f d s f s s df f df v d f df f f
d v df d s f f sdf d d f fs ff vfv d f v s vv d dv df ff ff fs f d f f s s d d d vf f f s f dd d s f
o d f f s d df ssf f s f f f f f ffv f f f s ff v ff df d f d f fd f s f f f s f d s sf f f s f dv s d o
60 o vf f f ff v s sv fv f f df f fv d f o sf fs f s vd v f s f ff ff d f s s d vfv f d s v sv f s
fv s d d f f s df df ff f f d fs s fs vs fd v fv s s f vf df ff v f s f d vfd f f s f df f d
sf f f d f v s ss f f dv f f f d v f d ss ff f ffd d d f d f df f f f f s d d fs f f v s s df f s
fd f f s f f d ff f fs df ff vv f v vs ff f d f sd f d ds f v d f v s f f fs ff f sv s f d s
vf d v v f f s ff v s ff fd fv f f vf f f s f s v f d f f fd ff f s f d sf s f sv d f d d f
55 fv s f f f svd d v o f d d d v ff f s d ff d f fs f f f d f f f v s fd fs f f d v d f f f o
ff d f f d fs fs d f dd dd sf f d v d ff d f df f s df f f f f f s d ff s f o df d f f v
f d df d f s vf f s dd dv f s s f s f sd s d dd s f dff f d f d s f s ff f df f s f ff
ff d f d f fs sf f f ff ff v f vv v s ff s fd f f d df df f f sf f f ff v dd d v d v
o f d f f f f ff d d ff f f d v dv s s ff f dfv v f df df v f s f fd d f f d v d f ff
50 d f f fd f fff d f d ff fv s ff s d d f v dfv v f fv s o v f s s d ff f f f f f d f
o f f o d f f ff d of df f v v ff v s f f vf sf f fff v ff s s f df f f ff d ff d
f f ff f vsv f f d fv v v f f s v f f s d sf s ffd d f f s f ff fd ff f d d d
f f d
f f ff f f o ff s f s f f s s d d d f d f f ffv d f d s f f fd f df f f f s
f f df o f ff d d f d f f s o f v v ff s d f f vf f d v ff s d v ff df df v d
45 f s sv o ds f d s d s vff d v v dd s s f o s f f d f d s d fff sf f d f s
f f sf f fs f s o f v ff f f s f sf s s v f fd d o o v s d d ff f sd d d d
f sf s o ds f f f d v sf o d s fd f s v f vf f d o df s v ds f s f d d d
v v f d o ss f v d f f f d o v ff s v f df f df f d s d f s f sv d d d
f f s v ss d f f f s fd d fs fs s s o f ff d f f s v f s d s s f d d
40 f d s o ss d f f fv d v f v dv d f f o v fs d f f v f d fs s s s d s
v f s f vs df f s v v f f s ss d vf df f f f v s d f s ff s s s s
Model Order ()

o f s o vs o d ff v f f s s f v s f o sf f f f f s d f s o s s f s
f f s dsf f f d v f f v v d d v fv d v f f f v s v s o s s d f
o d s fd f d o f f vf f v d s d v f df f o f s s f df s s v s
35 d f d vd f f o f s f f f f f d d f o df d f v f d sf s s s f s
s f s dsf f o f f vo d v s d s f ff d d f f f f f s s d s
o f s fs f f o f fv s d o d f f ff f f df f v s s d f f
f d fv o vf f f s f v s d o d v d d f f f fs d f s d f f
o f v ff d o d f o v s ff d d s s f f f f s vs s d ff
30 v d fd d o f v d d f s d f v d f f d f f ff s d v f
f d fs v f f v vf v s s f d f f d f v f s f s s f o
o d fs f f f v o v s f v f d fd f f v ds s d f f
o s fd f f v s s f s d d f d f f o s f f d s s d f
o v o s f f f s f v v s f s f f o o d f f s d f o
25 o s f s d o d f f v f d f d f s d f f v d d f
v v s f f v f f f o v f s s d f f s d s s v
d o v f o v o o f f o v d d f sf v d f
fs d f o ds d f o ff f d f o df d s v
s s s v o ds f s f d s d f sf v s d d
20 v s v o v f o f f f v d o f f s s f v
s v s o s o s f f s d f o d s v f f
v v d o f o v v o f d o f v f f o
f v d ff o v f f d o o f d f o
v s f f o s f o d f f v f ff
15 v f f v f s f d d f d o f f
o v f v o v f v s f f f d
f o o v s f f d f d d d
o f f o f f f f o o f f
f d o o f f d v o f
10 f f o f f f v d f
f f o o o v o f
o f o o f o o o
o o o f f o
o o o o o
5 o o f d
o o o
o f
o o
0
35 40 45 50 55 60 65 70 75
Frequency (Hz)

Figure 3: Stabilization diagram for Body data, analyzed with LSCE.

The first data set is a multiple-input multiple-output (MIMO) data set from the body of a car. It has 2 inputs
and 264 measurement points all over the car body, leading to 528 FRFs. The parameter estimation was done
with both a time domain method (LSCE) and a frequency domain method (PolyMAX) for the frequency
band 3575Hz. Both stabilization diagrams were created to a model size of 64.
The second data set is an output-only data set, from an in-flight test of an aircraft. During the measurements,
no artificial excitation was applied. The plane was just excited by atmospheric turbulence, which cannot be
measured, but is assumed to be white noise. It has 3 reference points and 7 outputs from various points all
over the airplane, leading to 21 spectra. The parameter estimation was done again with both an operational
time domain method Balanced Realization, BR, which is a subspace identification method [4][24] and
a frequency domain method (Operational PolyMAX) for the frequency band 0.502.26Hz. This may sound
strange for an in-flight data set, but the data was rescaled due to a non-disclosure agreement with the data
 InFlight BR

60 s o s s s s s v v v d ff s o s fs v d df vf f v s v f vs v ss f o f sf f f ff f f f v f s ff v
s o o v s s v s f s f fs s f f f d dff f f o v v v fsd f df v fv v v f s f ff f ff v f f v v
s d v v s s s v v f s f vf v f f f f ff f f o f v v f df f ff v f v v f f v fv f s v ff f f v
s s v v s s v s v v s v s o v v f f ff f ff f f f f f f ff f f f f f f ff vs v v v s s s s f d v f
s s v s v s s s s s s v s fs s dv v ff ff o f d f f f o vf v o f f f vs s d v s v s s f s v vf
55 d o f f v s d v f f s v s fv v vv f f f f d f o f f f ff fv f f f v f vv d v v v f v s d f f vf
f fo o v s f v o f v f s v vo dv f o ff f f o f f f ff ff f s f s v s vv s v s s v s s s s v fv
f o v v s s v v s v v s s v o f vf f o vf f f f f ff f f fd v s d vs v f s s s v s v d v fs
d o v v s v f o o v f s v fv f f f o f f f f f f ff f f o f f v sv v d v s v v v f f f f
o o f v s f f o o f o f v ff o o ff fv v ff f ff vf f o vf f v sv v v s v s s v v s v v
50 s v s s s vf v f s v f f f f f ff f o f f f f fv f f f f v vv f s s v s s v f f f v
f f o v o o o o f o v v d o f d f sf f f s f fv vf f o s f f vv f v v d v s s v v vf v
f v v fo o fo f f ff f f f ff f f v v f f v f f v f vv f v f f v v f f f v v
f o f v f v f vd f v of ff f f f f o f s v vf fv v o f ff sv f s v s s s v f v f f v
f v v f v v fv v v f o f o f f f f f s f o fs f v s s vf sv v f v d v s s f f o v
45 f v o s s f o sd s s f v f o f v d f v v f fs v fv f ff vv f f f d v v s f f f
v v s s v f sf s v f vf o f f f o f f fd f f v f s s s s f f f v v v s f d f
o f v o o f vf s v f d f v f ff f f f ff o v v f f v s f f v v f f v vf d f
v v f s v vs s s s v v vf s f v f f f v f o f s f v ff d f f f v d f f
o o v o v f dv v s f f f ff d f f f f f of ff v v v v v f f ff f f f
40 s v s v vv v s f vf d f f fo v s f f s f f f f v fs f f v f f s
v d v f fd d v f f o o ff o d f f o f f s fv o d f
d f s v s f f
ss v f vf s s f f o f f o d fv f o f f f vf f f v fv v f f o o
v f v f f d f d f o fo f f f o f f f d ff d f d ff f o f o
sd v s vv v v d v fs o f f o d o ff d f v f fv v s v f f
Model Order ()

35 vv v v vo d f f v ff f f f o f f f f v f ff f f df o f
vo f f f o f v d f vf o s f o f o f df v f v v s v f v f
v fo o v f ff f f ff o v s v f f v v f fd f f v v o
s fo o f v f f f f o f f o o o f f d o o v f f f o
v f fs s v v v f f f f s o vf f sf s f ff f f f o
30 v f o v o v s s o o f f f f o vf f s v o s fv s v f
o o ff o o v o s o o f o f o ff o f v v f o v s s s
o d d s vf v f v o f f f f o ff f v d ff f v f o
o o v ff f o v o f s o f f ff f v f d f v f
o o o o o s f v o f s f f f f o f v f v f f d o f
25 v vf v v of o f o f f v ff f d d v s v o
o s ff s s s o f f f f f f ff d f f v f
f f f f f o s o d ff f f f f f f f
v v s fv v f v v o o o v f v f f f v d
o o v o o o o o f v f v f v s f f v v
20 s o v f v f o v o f f o f s v
f o s f f f f s o f f f f f f o
o o f o o f f o f o f o f f v f o
v o o o f o o v d o f f v f
o f o o fo f ff v s f v f o
15 o o f ff o f f f f f f
s f f f f o s s s v v v
v d o o f f d d f f s v
v o o f f v o f o v o
o v o f o f o s v o
10 o o v o v o f v o
o o f o o ff v
s s d f fv s
s f o o s v
v o o s s
5 o o o s v
f o s f
s o d
v o
o
0
0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
Frequency (Hz)

Figure 4: Stabilization diagram for InFlight data, analyzed with BR.

Body PolyMAX InFlight PolyMAX

s s s 80 sd s s s s ssv sf d sfdff d dfod s s vvf fff sss svf dff o

s
s s
s ss of fff ff ss
f s
v
o ss d s ss s
vf
ss vvss s
sss
sf sss
ffff ss ss ss ss ss ss df ss vv vs ff ff sf f df vs ss
s
s s
s s fv
ff sss sss ss sss d ss vvs s sss
ffffd sss
s sss sss ss ss df d d vv vs ss df df
df df
df vf of df vv ss o ff vs vv d
d o
60 s
s s
s ss
s vf df ss d s ss dsf sss fffsss ss ss ss svs vf ss ds d d ss ss ss ff sf sf vf fv sf ss ff ss d sf
s s
ss sss ss sss
ff vs ss d s ss s s ssf s
sss sfffds v
sssf ss sss sss s f ff vo fd v
ss f s f ss sss s sff ss vds s ff ds ds sss d sf
70 ss sf ff ss sv ss ff ds ff ds sf fd sv ss
df of ss vf vf
s
s s
s s
ss vvf ss s ff s ss ss f ss ssf ds ffss ss ss ss vf d d ss vv vv ff ss ff ss ff odf vs ss ss o sf
s s vf sss sss vff s sffff sss sss sss
s d
vss fff s s f ss vs vv vf ss ff d o s sv ss
s
s s
s sss v v d ss sss ss ss s f f s d ss sf
50 s
s s
s d of sss d sf sss ss s vsf sss d ss ssvffff sss sss sss vss o s 60 ss ff ff vs sv sv ff ds of df o d
s d v ss ss ff
s s s o ss vs sss s
ss ss vfsss ss s
sff ss
s s s ss fo ss s ss ss s ss o d d ss ss vs vf df ss ss ss ss ss vf
ss
s
s
s
s ss ff s
f ss vsf ss ss
f
ss
s
sss fs fs sv
ss
svf fffsss sss sss vv ss o df vs vv vs dv vd ss d
d d s ss ss o
s
s s
s ss vsd sss vvf ss df ss fs dd
svs
s
svs sff sv dd ss vs vf of sv ss ff sf ss ss vv
40 s s ss ss d ss s
ss fs fss d
d ff fs fs ss ss 50
Model Order ()

Model Order ()

ss ss d s f ss
f o ss ff s ss ss sf vs
vvf d fs
fs ss ss ds o ss df ss sv vs vs ss df ss ss ss df
s
s s
ss ssf ov sss df ss vf svf f ds s
ss sss vsf ss ff ss ov vf df ss ff vs ss ss vf
s sss dd ss
ss ss
of of sfffsf
s
ds sss d fv o fsss d s f sss sss vff sv vd vs ff dd ds d
d d s ss ss ff
s sss of
f s
s ss d s v
f f sdf 40 sv ds vv oov ds d s ff sf ss ss of
30 sss sss d f fs
ov s d
ss
s d
sss d ss
s
sv vs ssdf s s
ss d s o
vs s ss ossv s s ss v sv ff vv vf sf sf df d d ss ss
f ss d f v ss vd ss ss f v ssvv
dfd fff sss sss vvv s
s d d vs v
v of ds ff ff ss ss
vf s
s df ossf sss d
d
d vs d s f o
sf
ssf sf d s ff 30 ss ff vs vv df dd df vf ss ss
dvf sss sff vs
s sss d
dv sss sd s ffv ssf
vsf ssv
fv
ss df vs vf fo vd ff oo ss ss o
20 sf v s v svs s ff d d o f vsv of v
v f f o oo ff dd df f ss ss
ss sff sss s dff d d vv v vof dff
o d
d s v f ss d
d ss d d
d vo vv 20 vs of of o d
d df ff vs ss v
o
d
v f v f ff o d dff ss o oo dd of ff ss ss
vfs fff vv
o o sv f vd ff ff
of ss o ss df ss vv
fof o df o
ss
10 f f f o of d
v f
d s
d sv
of o o o ff 10 v
o ff o ss vf
o oo fo ov o
o o
0 0
35 40 45 50 55 60 65 70 75 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
Frequency (Hz) Frequency (Hz)

Figure 5: Stabilization diagrams for Body and InFlight data, both analyzed with PolyMAX.

provider. The BR diagram was created up to a model size of 60, whereas the PolyMAX diagram was created
up to model size 80.

6.1 Observations

A first observation of the stabilization diagrams showed that the LSCE/BR stabilization diagrams the
former standard were quite complicated, whereas the PolyMAX diagrams were quite clear, showing clear
poles and only clear poles throughout the whole frequency band. This resulted in large differences in
the number of selected poles by the different test subjects. Figure 6 shows the distribution of the number of
selected poles of the 4 stabilization diagrams. The dark bar represents the bar in which the AASD results are
included. A first observation is that in general, the AASD tool is always in the high end of this spectrum,
suggesting an overestimate of the number of poles.
Also the time of the assessment was measured. In general the LSCE/BR tasks took about twice as much time
as the PolyMAX tasks. Next to that, the experts spent about twice as much time in the assessment as the
novices. The novices were quite overwhelmed by the complexity of the LSCE/BR diagrams, in which they
quickly gave up. AASD assessment takes only a few seconds (<10sec).
 Body LSCE Body PolyMAX InFlight BR InFlight PolyMAX
3 3 3 3

2 2 2 2

1 1 1 1

0 0 0 0
5 7 9 11 13 15 17 19 21 23 10 12 14 16 18 20 6 8 10121416182022242628303234 7 9 11 13 15 17
# Poles # Poles # Poles # Poles

Figure 6: Distribution of the number of selected poles per user for the 4 different stabilization diagrams.

Body LSCE Body PolyMAX

8 8
User (nr)

User (nr)
6 6
4 4
2 2
0 0
35 40 45 50 55 60 65 70 75 35 40 45 50 55 60 65 70 75
Frequency (Hz) Frequency (Hz)

Figure 7: Body LSCE and PolyMAX pole selection per user.

Body LSCE Body PolyMAX

0.04 0.04

0.035 0.035

0.03 0.03

0.025 0.025
Damping ()

Damping ()

0.02 0.02

0.015 0.015

0.01 0.01

0.005 0.005

0 0
35 40 45 50 55 60 65 70 75 35 40 45 50 55 60 65 70 75
Frequency (Hz) Frequency (Hz)

Figure 8: Body LSCE and PolyMAX pole selection, frequency vs. damping.

Figure 7 shows a frequency spectrum of the pole selection of all the test subjects on the two body stabilization
diagrams. Users 14 are the novices, where users 58 are the experts. The vertical dotted lines show the
selection made by the AASD tool. The LSCE diagram clearly shows that the novices had a hard time in
the 4560Hz band. A majority of the experts select poles at about the same frequencies as AASD, but the
alignment is not that nice. AASD selects 5 poles that have not been selected by any of the test subject but
almost all of the other poles are also selected by a majority of the experts. In one case at about 42.5Hz, the
test subjects have selected one pole, where AASD has selected two. This means that the users feel that two
branches at a lower order have stabilized into one pole, whereas AASD has found enough evidence to mark
them as separate poles. Finally, at 65.6Hz a majority of the experts have selected a pole that was missed by
AASD and all the novices.
In the PolyMAX diagram, AASD has selected two poles that no (or hardly any) subject has selected. There
are two poles that were selected by only two test subjects. All other poles were also selected by a big
majority.
Next, the damping of the selected poles is compared, and presented in figure 8. The crosses represent the
AASD selection. A cluster of dots represents consensus over the selection by the different test subjects,
whereas a scatter over the damping would indicate indecisiveness. In general the users have more consensus
with respect to damping in the PolyMAX diagram than in the LSCE diagram. AASD largely agrees with
this consensus. The differences are explained by the sparse number of selections on the 4560Hz band.
Remarkable is the improvement for both novices and experts in that band in the PolyMAX plot.
 InFlight BR InFlight PolyMAX
User (nr) 8 8

User (nr)
6 6
4 4
2 2
0 0
0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
Frequency (Hz) Frequency (Hz)

Figure 9: InFlight BR and PolyMAX pole selection per user.

InFlight BR InFlight PolyMAX

0.06 0.06

0.05 0.05

0.04 0.04
Damping ()

Damping ()
0.03 0.03

0.02 0.02

0.01 0.01

0 0
0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
Frequency (Hz) Frequency (Hz)

Figure 10: InFlight BR and PolyMAX pole selection, frequency vs. damping.

Figure 9 shows the frequency spectrum again of the pole selection of all the test subjects but now on the two
in-flight stabilization diagrams. Users 14 are the novices again, and users 58 are the experts. The vertical
dotted lines show the selection made by the AASD tool. Most remarkable is the trouble that the novices had
in the 1.01.6Hz band, where the experts agree with AASD. Overall, a majority of the experts select poles
at the same frequencies as AASD. Next, there are 3 AASD poles that were hardly selected by any of the test
subjects, and four poles that were only selected by two subjects.
In the PolyMAX diagram, there is a big consensus over almost all the poles. This time there is no significant
difference between novices and experts. There are two poles that are next to AASD only selected by two
expert test subjects. Next there are three poles that have not been selected by AASD, but only by two experts.
Figure 10 shows the damping of the selected poles. The crosses represent again the AASD selection. In
general the users have again more consensus with respect to damping in the PolyMAX diagram than in the
BR diagram. In analogy to the body stabilization diagrams, the AASD poles fit also nicely in the consensus
clusters of the test subjects. The exceptions in the BR diagram and in the PolyMAX diagram correspond to
the previously discussed poles that were hardly selected by the test subjects.

6.2 Using Validation Criteria

Typically the differences between measured and synthesized FRFs or spectra are used to asses the validity of
a modal model (see Section 5). As an example, this validation criterion is applied to the subspace identifica-
tion results of the in-flight data. In Figure 11, the measured cross spectra are compared with the cross spectra
synthesized from the identified modal parameters. The information of 21 (7 3) spectra is compressed by
using their 3 principal components (also called CMIF in modal analysis). The left figure is the result of a
novice, selecting only a few modes; the right one represents the AASD results, which are very close to the
results of the expert users. The main conclusion is that AASD is able to successfully extract the relevant
dynamic information from the in-flight data.
This criterion has to be applied with some care as it is well known that the fit always improves when more
parameters are included in the model. Figure 12 shows the correlation between the 3 CMIFs computed from
the measurement data and the CMIFs computed from the identified modal model. As it is typical that the
correlation increases with an increasing number of modes, there is only a marginal improvement after 20
modes. The fact that the correlation is not uniformly increasing with an increasing number of modes can be
explained by the human factor: some subjects reached a better correlation with less modes due to a better
selection.

0 0
10 10

1 1
10 10
CMIF []

CMIF []
2 2
10 10

3 3
10 10

4 4
10 10
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
f [Hz] f [Hz]

Figure 11: CMIF plots of in-flight BR data set for worst novice fit (left) and AASD close to best fit (right).

100
CMIF Correlation [%]

90

80

70

60
5 10 15 20 25 30 35
Number of modes []

Figure 12: CMIF correlation plot for increasing number of modes for in-flight BR data set.

For a discussion on the Modal Phase Collinearity (MPC) and Modal Phase Deviation (MPD) criteria, the
pole selection is used that was retrieved from the body data set using the LSCE method, as well as the
InFlight PolyMAX diagram. First, a set of poles was constructed on which there seemed to be consensus
based on observation. This can be visually validated by observing figure 3. Next, all the MPC and MPD
values for those poles were calculated. In order to compare the AASD tool with the novices and the experts,
average values were taken for each category. If a test subject did not select a pole, this was not included in
the average. The results are presented in table 1. Note that high MPC values as well as low MPD values
indicate physical poles.
AASD has a better MPC than the experts for 5 out of the 9 consensus poles, and has a better MPC than the
novices for 6 out of 9 consensus poles. AASD has 2 poles for which it scores remarkably low. In the case
of pole 5 at 42.4Hz, the diagram of figure 8 shows us that AASD has considered 2 poles there, where the
experts and novices have considered only 1. AASD has probably selected 2 poles at a low model order that
reflect the same physical phenomenon, but have not yet stabilized. The score of the experts and the novices
show that there is evidence for the existence of a physical pole at that frequency. The other remarkable pole
is pole 19 at 64.5Hz. Besides a low MPC for AASD, the novices and also the experts only reach a MPC
of about 60%. Inspecting the mode shape animation showed that this mode is indeed a complex mode but
still physically meaningful, where in general, true physical modes not do have complex shapes. Note that
trimmed car bodies sometimes have complex mode shapes due to local damping effects.
There remains one remarkable MPC value for the novices: they completely missed pole 15 at 56.5Hz, which
was situated at the edge of the difficult frequency band 4560Hz. The MPD values in table 1 show exactly the
same result as the MPC values. The high values for AASD for poles 5 and 19 reflect the previous discussion.
Table 2 shows the MPC and MPD values for the InFlight data set and the PolyMAX stabilization diagram.
 MPC [%] MPD [deg]
pole # Freq[Hz] Damp[%] AASD Avg.Nov. Avg.Exp. AASD Avg.Nov. Avg.Exp.
4 40.0 2.46% 93 86 88 17 23 21
5 42.4 1.85% 32 77 86 62 33 24
14 55.5 1.95% 81 61 71 29 46 38
15 56.5 0.50% 81 - 70 29 - 38
19 64.5 0.71% 37 62 62 61 43 45
20 66.6 1.71% 80 79 81 34 39 31
21 70.3 1.22% 98 82 96 10 29 13
22 71.1 1.27% 95 89 88 14 22 22
23 72.6 1.65% 70 76 77 40 36 33

Table 1: MPC and MPD: Body with LSCE, split into AASD, Average Novices, and Average Experts.

MPC [%] MPD [deg]

pole # Freq[Hz] Damp[%] AASD Avg.Nov. Avg.Exp. AASD Avg.Nov. Avg.Exp.
1 0.71 5.51% 81 99 100 34 6 4
3 0.78 2.91% 68 75 72 37 32 34
4 0.88 2.35% 80 80 83 27 28 25
5 0.95 0.83% 95 96 97 15 12 13
6 1.00 1.28% 95 89 92 19 26 26
7 1.24 2.20% 96 98 95 11 7 11
8 1.30 2.83% 94 - 90 20 - 25
9 1.58 1.19% 96 85 79 13 25 30
10 1.62 1.84% 98 99 99 10 7 6
11 1.68 4.52% 100 100 100 1 2 3

Table 2: MPC and MPD: InFlight with PolyMAX, split into AASD, Average Novices, and Average Experts.

A first observation is that the novices performed equally well as the experts, sometimes even slightly better,
even though they all missed pole 8 at 1.30Hz. The second observation is that AASD performs poorly on pole
1. The MPC is more or less acceptable, but the MPD is really bad, especially compared with the scores of
the experts and novices. However, on pole 6 and 9, AASD has a significantly better MPD value than both
experts and novices, and also the MPC is much better.
In general, it can be said that the novices score equally well on the PolyMAX diagram as the experts do.
Apart from pole 1, AASD also performs at or sometimes above the level of the experts.

7 Conclusions

One of the first observations was that AASD selects a rather high number of poles, as do the experts. It can
be concluded however, that this does not seem to impact the results, as pointed out by the validation criteria
CMIF, MPC, and MPD.
Next, the results showed that for difficult stabilization diagrams, acquired from parameter estimation methods
like LSCE or BR, the quality of the AASD selection meets those made by the experts rather than those made
by the novices. For clean PolyMAX stabilization diagrams, the tool and also the novices(!) perform at least
equally well as the experts. This supports the statement that PolyMAX yields user-independent results.
Finally, it has to be said that computationally, AASD by far outperforms human users in the order of seconds
versus minutes. We therefore suggest that the AASD tool will have an advisory role in the modal analysis
chain, with the adagio no gain no pain. Both expert and novices could then start off with the AASD
selection, and make changes when they think this is needed, thereby saving valuable time. For applications
such as on-line flutter monitoring, this could also proof to be a valuable automation step.
Based on this research, we recommend as an improvement for AASD that it also uses validation criteria
to assess its own selection. The heuristic that complex modes do not tend to be physical modes could be
implemented using the MPC/MPD validation criteria. This could however also lead to an elimination of
complex modes, that still are true physical modes, like in the Body/LSCE case.
Also, depending on the expertise of the user the tool could be fine-tuned to user scenarios, such as:

Overestimation: focussing on recall (percentage of the total number of physical poles found)

Underestimation: focussing on precision (percentage of selected poles that are also physical)

A mix of these two

At this moment, AASD tends to be tuned to overestimate, thereby focussed on not missing any poles, which
could be well suited for experts because they have the expertise to remove poles that they feel are not physical.

8 Acknowledgement

This work was carried out in the frame of the EC research project NMP2-CT-2003-501084 INMAR (In-
telligent Materials for Active Noise Reduction); see also http://www.lbf.fhg.de/inmar. The support of the EC
is gratefully acknowledged.

References

[1] Van der Auweraer, H. and Peeters, B. Discriminating physical poles from mathematical poles in high
order systems: Use and automation of the stabilization diagram. Proc. of IMTC2004, the 21th IEEE
Instrumentation and Measurement Technology Conf., Como, Italy, pages 21932198, 2004.

[2] Lanslots, J. and Scionti, M. Automatic Assessment of Stabilization Diagrams. Technical Report, LMS
International, TST Research & Technology Development, Leuven, Belgium, 2002.

[3] Scionti, M., Lanslots, J., Goethals, I., Vecchio, A., Van der Auweraer, H., Peeters, B., and De Moor,
B. Tools to Improve Detection of Structural Changes from In-Flight Flutter Data. Proc. of the VIII Int.
Conf. on Recent Advances in Structural Dynamics, Southampton, U.K., 2003.

[4] Peeters, B. System Identification and Damage Detection in Civil Engineering. Ph.D. Thesis, Dep. of
Civil Eng., K.U. Leuven, Belgium, http://www.bwk.kuleuven.ac.be/bwm, 2000.

[5] Heylen, W., Lammens, S., and Sas, P. Modal Analysis Theory and Testing. Katholieke Universiteit
Leuven, Department of Mechanical Eng., PMA, Leuven (Heverlee), Belgium, 1997.

[6] Verboven, P., Parloo, E., Guillaume, P., and Van Overmeire, M. Autonomous structural health moni-
toring part i: modal parameter estimation and tracking. Mechanical Systems and Signal Processing,
16(4):637657, 2002.

[7] Verboven, P., Guillaume, P., Cauberghe, B., Parloo, E., and Vanlanduit, S. Stabilization charts and
uncertainty bounds for frequency-domain linear least squares estimators,. Proc. of IMAC 21, the Int.
Modal Analysis Conf., Kissimmee, Florida, USA, pages 637657, 2003.

[8] Chhipwadia, K.S., Zimmerman, D.C., and James III, G.H. Evolving autonomous modal parameter
estimation. Proc. of IMAC 17, the Int. Modal Analysis Conf., Kissimmee, Florida, USA, pages 819
825, 1999.
 [9] Goethals, I. and De Moor, B. Subspace identification combined with new mode selection techniques for
modal analysis of an airplane. Proc. of SYSID-2003, the 13th IFAC Symposium on System Identification,
Rotterdam, The Netherlands, 2003.

[10] Pappa, R.S., James III, G.H., and Zimmerman, D.C. Autonomous model identification of the space
shuttle tail rudder. Journal of Spacecraft and Rockets, 35(2):163169, 1998.

[11] Juang, J.-N. and Pappa, R. An eigensystem realization algorithm for modal parameter identification
and reduction. Journal of Guidance, Control and Dynamics, 8(5):620627, 1985.

[12] Verboven, P., Parloo, E., Guillaume, P., and Van Overmeire, M. Autonomous modal parameter identi-
fication based on a statistical frequency-domain maximum likelihood approach. Proc. of IMAC 19, the
Int. Modal Analysis Conf., Kissimmee, Florida, USA, pages 15111517, 2001.

[13] Vanlanduit, S., Verboven, P., Schoukens, J., and Guillaume, P. An automatic frequency domain modal
parameter estimation algorithm. Proc. of Int. Conf. on Structural System Identification, Kassel, Ger-
many, pages 637646, 2001.

[14] Goethals, I. and De Moor, B. Model reduction and energy analysis as a tool to detect spurious modes.
Proc. ISMA 2002, the Int. Conf. on Noise and Vibration Eng., Leuven, Belgium, pages 13071314,
2002.

[15] Lanslots, J., Scionti, M., and Vecchio, A. Fuzzy Clustering Techniques to Automatically Assess Stabi-
lization Diagrams. Proc. of the 7th Int. Conf. on the Application of Artificial Intelligence to Civil and
Structural Eng., Egmond-aan-Zee, The Netherlands, paper 46, 2003.

[16] Van Overschee, P. and De Moor, B. Subspace identification for linear systems: theory, implementation,
applications. Kluwer Academic Publishers, Dordrecht, The Netherlands, 1996.

[17] Leuridan, J. Some direct parameter model identification methods applicable for multiple input modal
analysis. Ph.D. Thesis, Univ. of Cincinnati, USA, 1984.

[18] Busturia, J. and Gimenez, J. Multi-excitation multi-response non-linear least squares algorithm. Proc.
of 10th ISMA, the Int. Conf. on Noise and Vibration Eng., Leuven, Belgium, 1985.

[19] Vold, H., Kundrat, J., Rocklin, T., and Russel, R. A multi-input modal parameter estimation algorithm
for mini-computers. SAE paper 820194, Trans. SAE, 91(1):815821, 1982.

[20] Guillaume, P., Verboven, P., and Vanlanduit, S. Frequency-domain maximum likelihood identification
of modal parameters with confidence intervals. Proc. ISMA 23, the Int. Conf. on Noise and Vibration
Eng., Leuven, Belgium, 1998.

[21] Van der Auweraer, H., Guillaume, P., Verboven, P., and Vanlanduit, S. Application of a fast-stabilizing
frequency domain parameter estimation method. ASME Journal of Dynamic Systems, Measurement,
and Control, 123(4):651658, 2001.

[22] Guillaume, P., Verboven, P., Vanlanduit, S., Van der Auweraer, H., and Peeters, B. A poly-reference
implementation of the least-squares complex frequency-domain estimator. Proc. of IMAC 21, the Int.
Modal Analysis Conf., Kissimmee, FL, USA, 2003.

[23] Peeters, B., Guillaume, P., Van der Auweraer, H., Cauberghe, B., Verboven, P., and Leuridan, J. Auto-
motive and aerospace applications of the polymax modal parameter estimation method. Proc. of IMAC
22, the Int. Modal Analysis Conf., Dearborn, MI, USA, 2004.

[24] Hermans, L. and Van der Auweraer, H. Modal testing and analysis of structures under operational
conditions: industrial applications. Mechanical Systems and Signal Processing, 13(2):193216, 1999.